arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-23 00:28:11 +00:00
Code Issues Releases Wiki Activity

Cleanup -m 9800 kernels to latest standard

Browse Source
This commit is contained in:
Jens Steube 2016-04-17 15:49:02 +02:00
parent 8d34f3f266
commit e6b9071b52
5 changed files with 477 additions and 642 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
OpenCL/m09700_a3.cl
View File

@ -611,14 +611,7 @@ static void m09700m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
// now the RC4 part // now the RC4 part
u32 key[4]; rc4_init_16 (rc4_key, digest);
key[0] = digest[0];
key[1] = digest[1];
key[2] = digest[2];
key[3] = digest[3];
rc4_init_16 (rc4_key, key);
u32 out[4]; u32 out[4];
@ -1040,14 +1033,7 @@ static void m09700s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
// now the RC4 part // now the RC4 part
u32 key[4]; rc4_init_16 (rc4_key, digest);
key[0] = digest[0];
key[1] = digest[1];
key[2] = digest[2];
key[3] = digest[3];
rc4_init_16 (rc4_key, key);
u32 out[4]; u32 out[4];

413
OpenCL/m09800_a0.cl
View File

@ -5,6 +5,9 @@
#define _OLDOFFICE34_ #define _OLDOFFICE34_
//too much register pressure
//#define NEW_SIMD_CODE
#include "include/constants.h" #include "include/constants.h"
#include "include/kernel_vendor.h" #include "include/kernel_vendor.h"
@ -18,9 +21,7 @@
#include "OpenCL/common.c" #include "OpenCL/common.c"
#include "include/rp_kernel.h" #include "include/rp_kernel.h"
#include "OpenCL/rp.c" #include "OpenCL/rp.c"
#include "OpenCL/simd.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
typedef struct typedef struct
{ {
@ -277,10 +278,6 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* base * base
*/ */
@ -290,14 +287,12 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 pw_buf0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4]; pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5]; pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6]; pw_buf1[2] = pws[gid].i[ 6];
@ -305,6 +300,14 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 pw_len = pws[gid].pw_len; const u32 pw_len = pws[gid].pw_len;
/**
* shared
*/
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* salt * salt
*/ */
@ -316,8 +319,6 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
const u32 salt_len = 16;
/** /**
* esalt * esalt
*/ */
@ -335,68 +336,42 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
u32 w0[4]; u32x w0[4] = { 0 };
u32x w1[4] = { 0 };
u32x w2[4] = { 0 };
u32x w3[4] = { 0 };
w0[0] = pw_buf0[0]; const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
w0[1] = pw_buf0[1];
w0[2] = pw_buf0[2];
w0[3] = pw_buf0[3];
u32 w1[4];
w1[0] = pw_buf1[0];
w1[1] = pw_buf1[1];
w1[2] = pw_buf1[2];
w1[3] = pw_buf1[3];
u32 w2[4];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
const u32x pw_salt_len = (out_len * 2) + salt_len;
append_0x80_2x4_VV (w0, w1, out_len); append_0x80_2x4_VV (w0, w1, out_len);
u32 w0_t[4]; /**
u32 w1_t[4]; * sha1
u32 w2_t[4]; */
u32 w3_t[4];
make_unicode (w0, w0_t, w1_t); make_unicode (w1, w2, w3);
make_unicode (w1, w2_t, w3_t); make_unicode (w0, w0, w1);
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len); const u32x pw_salt_len = (out_len * 2) + 16;
w0_t[0] = salt_buf[0]; w3[3] = pw_salt_len * 8;
w0_t[1] = salt_buf[1]; w3[2] = 0;
w0_t[2] = salt_buf[2]; w3[1] = swap32 (w2[1]);
w0_t[3] = salt_buf[3]; w3[0] = swap32 (w2[0]);
w1_t[0] = swap32 (w1_t[0]); w2[3] = swap32 (w1[3]);
w1_t[1] = swap32 (w1_t[1]); w2[2] = swap32 (w1[2]);
w1_t[2] = swap32 (w1_t[2]); w2[1] = swap32 (w1[1]);
w1_t[3] = swap32 (w1_t[3]); w2[0] = swap32 (w1[0]);
w2_t[0] = swap32 (w2_t[0]); w1[3] = swap32 (w0[3]);
w2_t[1] = swap32 (w2_t[1]); w1[2] = swap32 (w0[2]);
w2_t[2] = swap32 (w2_t[2]); w1[1] = swap32 (w0[1]);
w2_t[3] = swap32 (w2_t[3]); w1[0] = swap32 (w0[0]);
w3_t[0] = swap32 (w3_t[0]); w0[3] = salt_buf[3];
w3_t[1] = swap32 (w3_t[1]); w0[2] = salt_buf[2];
w3_t[2] = 0; w0[1] = salt_buf[1];
w3_t[3] = pw_salt_len * 8; w0[0] = salt_buf[0];
u32 digest[5]; u32 digest[5];
@ -406,24 +381,24 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
w0_t[0] = digest[0]; w0[0] = digest[0];
w0_t[1] = digest[1]; w0[1] = digest[1];
w0_t[2] = digest[2]; w0[2] = digest[2];
w0_t[3] = digest[3]; w0[3] = digest[3];
w1_t[0] = digest[4]; w1[0] = digest[4];
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0x80000000; w1[2] = 0x80000000;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (20 + 4) * 8; w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -431,44 +406,42 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
u32 key[4]; digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]);
key[0] = swap32 (digest[0]); digest[2] = swap32 (digest[2]);
key[1] = swap32 (digest[1]); digest[3] = swap32 (digest[3]);
key[2] = swap32 (digest[2]);
key[3] = swap32 (digest[3]);
if (version == 3) if (version == 3)
{ {
key[1] &= 0xff; digest[1] &= 0xff;
key[2] = 0; digest[2] = 0;
key[3] = 0; digest[3] = 0;
} }
rc4_init_16 (rc4_key, key); rc4_init_16 (rc4_key, digest);
u32 out[4]; u32 out[4];
u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out);
w0_t[0] = swap32 (out[0]); w0[0] = swap32 (out[0]);
w0_t[1] = swap32 (out[1]); w0[1] = swap32 (out[1]);
w0_t[2] = swap32 (out[2]); w0[2] = swap32 (out[2]);
w0_t[3] = swap32 (out[3]); w0[3] = swap32 (out[3]);
w1_t[0] = 0x80000000; w1[0] = 0x80000000;
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = 16 * 8; w3[3] = 16 * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -476,7 +449,7 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
digest[0] = swap32 (digest[0]); digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]); digest[1] = swap32 (digest[1]);
@ -485,12 +458,7 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
rc4_next_16 (rc4_key, 16, j, digest, out); rc4_next_16 (rc4_key, 16, j, digest, out);
const u32 r0 = out[0]; COMPARE_M_SIMD (out[0], out[1], out[2], out[3]);
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_M
} }
} }
@ -510,10 +478,6 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* base * base
*/ */
@ -523,14 +487,12 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 pw_buf0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4]; pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5]; pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6]; pw_buf1[2] = pws[gid].i[ 6];
@ -539,16 +501,12 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 pw_len = pws[gid].pw_len; const u32 pw_len = pws[gid].pw_len;
/** /**
* digest * shared
*/ */
const u32 search[4] = __local RC4_KEY rc4_keys[64];
{
digests_buf[digests_offset].digest_buf[DGST_R0], __local RC4_KEY *rc4_key = &rc4_keys[lid];
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/** /**
* salt * salt
@ -561,8 +519,6 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
const u32 salt_len = 16;
/** /**
* esalt * esalt
*/ */
@ -576,72 +532,58 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
encryptedVerifier[2] = oldoffice34_bufs[salt_pos].encryptedVerifier[2]; encryptedVerifier[2] = oldoffice34_bufs[salt_pos].encryptedVerifier[2];
encryptedVerifier[3] = oldoffice34_bufs[salt_pos].encryptedVerifier[3]; encryptedVerifier[3] = oldoffice34_bufs[salt_pos].encryptedVerifier[3];
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/** /**
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
u32 w0[4]; u32x w0[4] = { 0 };
u32x w1[4] = { 0 };
u32x w2[4] = { 0 };
u32x w3[4] = { 0 };
w0[0] = pw_buf0[0]; const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
w0[1] = pw_buf0[1];
w0[2] = pw_buf0[2];
w0[3] = pw_buf0[3];
u32 w1[4];
w1[0] = pw_buf1[0];
w1[1] = pw_buf1[1];
w1[2] = pw_buf1[2];
w1[3] = pw_buf1[3];
u32 w2[4];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
const u32x pw_salt_len = (out_len * 2) + salt_len;
append_0x80_2x4_VV (w0, w1, out_len); append_0x80_2x4_VV (w0, w1, out_len);
u32 w0_t[4]; /**
u32 w1_t[4]; * sha1
u32 w2_t[4]; */
u32 w3_t[4];
make_unicode (w0, w0_t, w1_t); make_unicode (w1, w2, w3);
make_unicode (w1, w2_t, w3_t); make_unicode (w0, w0, w1);
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len); const u32x pw_salt_len = (out_len * 2) + 16;
w0_t[0] = salt_buf[0]; w3[3] = pw_salt_len * 8;
w0_t[1] = salt_buf[1]; w3[2] = 0;
w0_t[2] = salt_buf[2]; w3[1] = swap32 (w2[1]);
w0_t[3] = salt_buf[3]; w3[0] = swap32 (w2[0]);
w1_t[0] = swap32 (w1_t[0]); w2[3] = swap32 (w1[3]);
w1_t[1] = swap32 (w1_t[1]); w2[2] = swap32 (w1[2]);
w1_t[2] = swap32 (w1_t[2]); w2[1] = swap32 (w1[1]);
w1_t[3] = swap32 (w1_t[3]); w2[0] = swap32 (w1[0]);
w2_t[0] = swap32 (w2_t[0]); w1[3] = swap32 (w0[3]);
w2_t[1] = swap32 (w2_t[1]); w1[2] = swap32 (w0[2]);
w2_t[2] = swap32 (w2_t[2]); w1[1] = swap32 (w0[1]);
w2_t[3] = swap32 (w2_t[3]); w1[0] = swap32 (w0[0]);
w3_t[0] = swap32 (w3_t[0]); w0[3] = salt_buf[3];
w3_t[1] = swap32 (w3_t[1]); w0[2] = salt_buf[2];
w3_t[2] = 0; w0[1] = salt_buf[1];
w3_t[3] = pw_salt_len * 8; w0[0] = salt_buf[0];
u32 digest[5]; u32 digest[5];
@ -651,24 +593,24 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
w0_t[0] = digest[0]; w0[0] = digest[0];
w0_t[1] = digest[1]; w0[1] = digest[1];
w0_t[2] = digest[2]; w0[2] = digest[2];
w0_t[3] = digest[3]; w0[3] = digest[3];
w1_t[0] = digest[4]; w1[0] = digest[4];
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0x80000000; w1[2] = 0x80000000;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (20 + 4) * 8; w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -676,44 +618,42 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
u32 key[4]; digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]);
key[0] = swap32 (digest[0]); digest[2] = swap32 (digest[2]);
key[1] = swap32 (digest[1]); digest[3] = swap32 (digest[3]);
key[2] = swap32 (digest[2]);
key[3] = swap32 (digest[3]);
if (version == 3) if (version == 3)
{ {
key[1] &= 0xff; digest[1] &= 0xff;
key[2] = 0; digest[2] = 0;
key[3] = 0; digest[3] = 0;
} }
rc4_init_16 (rc4_key, key); rc4_init_16 (rc4_key, digest);
u32 out[4]; u32 out[4];
u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out);
w0_t[0] = swap32 (out[0]); w0[0] = swap32 (out[0]);
w0_t[1] = swap32 (out[1]); w0[1] = swap32 (out[1]);
w0_t[2] = swap32 (out[2]); w0[2] = swap32 (out[2]);
w0_t[3] = swap32 (out[3]); w0[3] = swap32 (out[3]);
w1_t[0] = 0x80000000; w1[0] = 0x80000000;
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = 16 * 8; w3[3] = 16 * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -721,7 +661,7 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
digest[0] = swap32 (digest[0]); digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]); digest[1] = swap32 (digest[1]);
@ -730,12 +670,7 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
rc4_next_16 (rc4_key, 16, j, digest, out); rc4_next_16 (rc4_key, 16, j, digest, out);
const u32 r0 = out[0]; COMPARE_S_SIMD (out[0], out[1], out[2], out[3]);
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_S
} }
} }

583
OpenCL/m09800_a1.cl
View File

@ -5,6 +5,9 @@
#define _OLDOFFICE34_ #define _OLDOFFICE34_
//too much register pressure
//#define NEW_SIMD_CODE
#include "include/constants.h" #include "include/constants.h"
#include "include/kernel_vendor.h" #include "include/kernel_vendor.h"
@ -16,9 +19,7 @@
#include "include/kernel_functions.c" #include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c" #include "OpenCL/types_ocl.c"
#include "OpenCL/common.c" #include "OpenCL/common.c"
#include "OpenCL/simd.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
typedef struct typedef struct
{ {
@ -275,10 +276,6 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* base * base
*/ */
@ -287,40 +284,27 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 wordl0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
wordl0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[0];
wordl0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[1];
wordl0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[2];
wordl0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
u32 wordl1[4]; pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
wordl1[0] = pws[gid].i[ 4]; pw_buf1[3] = pws[gid].i[7];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
const u32 pw_l_len = pws[gid].pw_len; const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT) /**
{ * shared
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len); */
}
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* salt * salt
@ -333,8 +317,6 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
const u32 salt_len = 16;
/** /**
* esalt * esalt
*/ */
@ -352,103 +334,92 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
const u32 pw_r_len = combs_buf[il_pos].pw_len; const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
const u32 pw_salt_len = (pw_len * 2) + salt_len; /**
* concat password candidate
*/
u32 wordr0[4]; u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordr0[0] = combs_buf[il_pos].i[0]; wordl0[0] = pw_buf0[0];
wordr0[1] = combs_buf[il_pos].i[1]; wordl0[1] = pw_buf0[1];
wordr0[2] = combs_buf[il_pos].i[2]; wordl0[2] = pw_buf0[2];
wordr0[3] = combs_buf[il_pos].i[3]; wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32 wordr1[4]; u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr1[0] = combs_buf[il_pos].i[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr1[1] = combs_buf[il_pos].i[5]; wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr1[2] = combs_buf[il_pos].i[6]; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr1[3] = combs_buf[il_pos].i[7]; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
u32 wordr2[4]; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr2[0] = 0; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr2[1] = 0;
wordr2[2] = 0;
wordr2[3] = 0;
u32 wordr3[4];
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{ {
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len); switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
} }
u32 w0[4]; u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0]; w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
u32 w2[4]; /**
* sha1
*/
w2[0] = wordl2[0] | wordr2[0]; make_unicode (w1, w2, w3);
w2[1] = wordl2[1] | wordr2[1]; make_unicode (w0, w0, w1);
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
u32 w3[4]; const u32x pw_salt_len = (pw_len * 2) + 16;
w3[0] = wordl3[0] | wordr3[0]; w3[3] = pw_salt_len * 8;
w3[1] = wordl3[1] | wordr3[1];
w3[2] = 0; w3[2] = 0;
w3[3] = 0; w3[1] = swap32 (w2[1]);
w3[0] = swap32 (w2[0]);
append_0x80_2x4 (w0, w1, pw_len); w2[3] = swap32 (w1[3]);
w2[2] = swap32 (w1[2]);
u32 w0_t[4]; w2[1] = swap32 (w1[1]);
u32 w1_t[4]; w2[0] = swap32 (w1[0]);
u32 w2_t[4]; w1[3] = swap32 (w0[3]);
u32 w3_t[4]; w1[2] = swap32 (w0[2]);
w1[1] = swap32 (w0[1]);
make_unicode (w0, w0_t, w1_t); w1[0] = swap32 (w0[0]);
make_unicode (w1, w2_t, w3_t); w0[3] = salt_buf[3];
w0[2] = salt_buf[2];
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len); w0[1] = salt_buf[1];
w0[0] = salt_buf[0];
w0_t[0] = salt_buf[0];
w0_t[1] = salt_buf[1];
w0_t[2] = salt_buf[2];
w0_t[3] = salt_buf[3];
w1_t[0] = swap32 (w1_t[0]);
w1_t[1] = swap32 (w1_t[1]);
w1_t[2] = swap32 (w1_t[2]);
w1_t[3] = swap32 (w1_t[3]);
w2_t[0] = swap32 (w2_t[0]);
w2_t[1] = swap32 (w2_t[1]);
w2_t[2] = swap32 (w2_t[2]);
w2_t[3] = swap32 (w2_t[3]);
w3_t[0] = swap32 (w3_t[0]);
w3_t[1] = swap32 (w3_t[1]);
w3_t[2] = 0;
w3_t[3] = pw_salt_len * 8;
u32 digest[5]; u32 digest[5];
@ -458,24 +429,24 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
w0_t[0] = digest[0]; w0[0] = digest[0];
w0_t[1] = digest[1]; w0[1] = digest[1];
w0_t[2] = digest[2]; w0[2] = digest[2];
w0_t[3] = digest[3]; w0[3] = digest[3];
w1_t[0] = digest[4]; w1[0] = digest[4];
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0x80000000; w1[2] = 0x80000000;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (20 + 4) * 8; w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -483,44 +454,42 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
u32 key[4]; digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]);
key[0] = swap32 (digest[0]); digest[2] = swap32 (digest[2]);
key[1] = swap32 (digest[1]); digest[3] = swap32 (digest[3]);
key[2] = swap32 (digest[2]);
key[3] = swap32 (digest[3]);
if (version == 3) if (version == 3)
{ {
key[1] &= 0xff; digest[1] &= 0xff;
key[2] = 0; digest[2] = 0;
key[3] = 0; digest[3] = 0;
} }
rc4_init_16 (rc4_key, key); rc4_init_16 (rc4_key, digest);
u32 out[4]; u32 out[4];
u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out);
w0_t[0] = swap32 (out[0]); w0[0] = swap32 (out[0]);
w0_t[1] = swap32 (out[1]); w0[1] = swap32 (out[1]);
w0_t[2] = swap32 (out[2]); w0[2] = swap32 (out[2]);
w0_t[3] = swap32 (out[3]); w0[3] = swap32 (out[3]);
w1_t[0] = 0x80000000; w1[0] = 0x80000000;
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = 16 * 8; w3[3] = 16 * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -528,7 +497,7 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
digest[0] = swap32 (digest[0]); digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]); digest[1] = swap32 (digest[1]);
@ -537,12 +506,7 @@ __kernel void m09800_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
rc4_next_16 (rc4_key, 16, j, digest, out); rc4_next_16 (rc4_key, 16, j, digest, out);
const u32 r0 = out[0]; COMPARE_M_SIMD (out[0], out[1], out[2], out[3]);
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_M
} }
} }
@ -562,10 +526,6 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* base * base
*/ */
@ -574,52 +534,27 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 wordl0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
wordl0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[0];
wordl0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[1];
wordl0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[2];
wordl0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
u32 wordl1[4]; pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
wordl1[0] = pws[gid].i[ 4]; pw_buf1[3] = pws[gid].i[7];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
const u32 pw_l_len = pws[gid].pw_len; const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* digest * shared
*/ */
const u32 search[4] = __local RC4_KEY rc4_keys[64];
{
digests_buf[digests_offset].digest_buf[DGST_R0], __local RC4_KEY *rc4_key = &rc4_keys[lid];
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/** /**
* salt * salt
@ -632,8 +567,6 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
const u32 salt_len = 16;
/** /**
* esalt * esalt
*/ */
@ -647,107 +580,108 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
encryptedVerifier[2] = oldoffice34_bufs[salt_pos].encryptedVerifier[2]; encryptedVerifier[2] = oldoffice34_bufs[salt_pos].encryptedVerifier[2];
encryptedVerifier[3] = oldoffice34_bufs[salt_pos].encryptedVerifier[3]; encryptedVerifier[3] = oldoffice34_bufs[salt_pos].encryptedVerifier[3];
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/** /**
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
const u32 pw_r_len = combs_buf[il_pos].pw_len; const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
const u32 pw_salt_len = (pw_len * 2) + salt_len; /**
* concat password candidate
*/
u32 wordr0[4]; u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordr0[0] = combs_buf[il_pos].i[0]; wordl0[0] = pw_buf0[0];
wordr0[1] = combs_buf[il_pos].i[1]; wordl0[1] = pw_buf0[1];
wordr0[2] = combs_buf[il_pos].i[2]; wordl0[2] = pw_buf0[2];
wordr0[3] = combs_buf[il_pos].i[3]; wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32 wordr1[4]; u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr1[0] = combs_buf[il_pos].i[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr1[1] = combs_buf[il_pos].i[5]; wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr1[2] = combs_buf[il_pos].i[6]; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr1[3] = combs_buf[il_pos].i[7]; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
u32 wordr2[4]; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr2[0] = 0; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr2[1] = 0;
wordr2[2] = 0;
wordr2[3] = 0;
u32 wordr3[4];
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{ {
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len); switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
} }
u32 w0[4]; u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0]; w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
u32 w2[4]; /**
* sha1
*/
w2[0] = wordl2[0] | wordr2[0]; make_unicode (w1, w2, w3);
w2[1] = wordl2[1] | wordr2[1]; make_unicode (w0, w0, w1);
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
u32 w3[4]; const u32x pw_salt_len = (pw_len * 2) + 16;
w3[0] = wordl3[0] | wordr3[0]; w3[3] = pw_salt_len * 8;
w3[1] = wordl3[1] | wordr3[1];
w3[2] = 0; w3[2] = 0;
w3[3] = 0; w3[1] = swap32 (w2[1]);
w3[0] = swap32 (w2[0]);
append_0x80_2x4 (w0, w1, pw_len); w2[3] = swap32 (w1[3]);
w2[2] = swap32 (w1[2]);
u32 w0_t[4]; w2[1] = swap32 (w1[1]);
u32 w1_t[4]; w2[0] = swap32 (w1[0]);
u32 w2_t[4]; w1[3] = swap32 (w0[3]);
u32 w3_t[4]; w1[2] = swap32 (w0[2]);
w1[1] = swap32 (w0[1]);
make_unicode (w0, w0_t, w1_t); w1[0] = swap32 (w0[0]);
make_unicode (w1, w2_t, w3_t); w0[3] = salt_buf[3];
w0[2] = salt_buf[2];
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len); w0[1] = salt_buf[1];
w0[0] = salt_buf[0];
w0_t[0] = salt_buf[0];
w0_t[1] = salt_buf[1];
w0_t[2] = salt_buf[2];
w0_t[3] = salt_buf[3];
w1_t[0] = swap32 (w1_t[0]);
w1_t[1] = swap32 (w1_t[1]);
w1_t[2] = swap32 (w1_t[2]);
w1_t[3] = swap32 (w1_t[3]);
w2_t[0] = swap32 (w2_t[0]);
w2_t[1] = swap32 (w2_t[1]);
w2_t[2] = swap32 (w2_t[2]);
w2_t[3] = swap32 (w2_t[3]);
w3_t[0] = swap32 (w3_t[0]);
w3_t[1] = swap32 (w3_t[1]);
w3_t[2] = 0;
w3_t[3] = pw_salt_len * 8;
u32 digest[5]; u32 digest[5];
@ -757,24 +691,24 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
w0_t[0] = digest[0]; w0[0] = digest[0];
w0_t[1] = digest[1]; w0[1] = digest[1];
w0_t[2] = digest[2]; w0[2] = digest[2];
w0_t[3] = digest[3]; w0[3] = digest[3];
w1_t[0] = digest[4]; w1[0] = digest[4];
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0x80000000; w1[2] = 0x80000000;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = (20 + 4) * 8; w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -782,44 +716,42 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
u32 key[4]; digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]);
key[0] = swap32 (digest[0]); digest[2] = swap32 (digest[2]);
key[1] = swap32 (digest[1]); digest[3] = swap32 (digest[3]);
key[2] = swap32 (digest[2]);
key[3] = swap32 (digest[3]);
if (version == 3) if (version == 3)
{ {
key[1] &= 0xff; digest[1] &= 0xff;
key[2] = 0; digest[2] = 0;
key[3] = 0; digest[3] = 0;
} }
rc4_init_16 (rc4_key, key); rc4_init_16 (rc4_key, digest);
u32 out[4]; u32 out[4];
u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out);
w0_t[0] = swap32 (out[0]); w0[0] = swap32 (out[0]);
w0_t[1] = swap32 (out[1]); w0[1] = swap32 (out[1]);
w0_t[2] = swap32 (out[2]); w0[2] = swap32 (out[2]);
w0_t[3] = swap32 (out[3]); w0[3] = swap32 (out[3]);
w1_t[0] = 0x80000000; w1[0] = 0x80000000;
w1_t[1] = 0; w1[1] = 0;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 0; w3[2] = 0;
w3_t[3] = 16 * 8; w3[3] = 16 * 8;
digest[0] = SHA1M_A; digest[0] = SHA1M_A;
digest[1] = SHA1M_B; digest[1] = SHA1M_B;
@ -827,7 +759,7 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[3] = SHA1M_D; digest[3] = SHA1M_D;
digest[4] = SHA1M_E; digest[4] = SHA1M_E;
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0, w1, w2, w3, digest);
digest[0] = swap32 (digest[0]); digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]); digest[1] = swap32 (digest[1]);
@ -836,12 +768,7 @@ __kernel void m09800_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
rc4_next_16 (rc4_key, 16, j, digest, out); rc4_next_16 (rc4_key, 16, j, digest, out);
const u32 r0 = out[0]; COMPARE_S_SIMD (out[0], out[1], out[2], out[3]);
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_S
} }
} }

84
OpenCL/m09800_a3.cl
View File

@ -276,6 +276,10 @@ static void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
const u32 gid = get_global_id (0); const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
/**
* shared
*/
__local RC4_KEY *rc4_key = &rc4_keys[lid]; __local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
@ -289,10 +293,6 @@ static void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
const u32 salt_len = 16;
const u32 pw_salt_len = pw_len + salt_len;
/** /**
* esalt * esalt
*/ */
@ -318,6 +318,12 @@ static void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
const u32 w0lr = w0l | w0r; const u32 w0lr = w0l | w0r;
/**
* sha1
*/
const u32 pw_salt_len = pw_len + 16;
u32 w0_t[4]; u32 w0_t[4];
u32 w1_t[4]; u32 w1_t[4];
u32 w2_t[4]; u32 w2_t[4];
@ -375,21 +381,19 @@ static void m09800m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
u32 key[4]; digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]);
key[0] = swap32 (digest[0]); digest[2] = swap32 (digest[2]);
key[1] = swap32 (digest[1]); digest[3] = swap32 (digest[3]);
key[2] = swap32 (digest[2]);
key[3] = swap32 (digest[3]);
if (version == 3) if (version == 3)
{ {
key[1] &= 0xff; digest[1] &= 0xff;
key[2] = 0; digest[2] = 0;
key[3] = 0; digest[3] = 0;
} }
rc4_init_16 (rc4_key, key); rc4_init_16 (rc4_key, digest);
u32 out[4]; u32 out[4];
@ -440,19 +444,11 @@ static void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
const u32 gid = get_global_id (0); const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/** /**
* digest * shared
*/ */
const u32 search[4] = __local RC4_KEY *rc4_key = &rc4_keys[lid];
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/** /**
* salt * salt
@ -465,10 +461,6 @@ static void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
const u32 salt_len = 16;
const u32 pw_salt_len = pw_len + salt_len;
/** /**
* esalt * esalt
*/ */
@ -482,6 +474,18 @@ static void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
encryptedVerifier[2] = oldoffice34_bufs[salt_pos].encryptedVerifier[2]; encryptedVerifier[2] = oldoffice34_bufs[salt_pos].encryptedVerifier[2];
encryptedVerifier[3] = oldoffice34_bufs[salt_pos].encryptedVerifier[3]; encryptedVerifier[3] = oldoffice34_bufs[salt_pos].encryptedVerifier[3];
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/** /**
* loop * loop
*/ */
@ -494,6 +498,12 @@ static void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
const u32 w0lr = w0l | w0r; const u32 w0lr = w0l | w0r;
/**
* sha1
*/
const u32 pw_salt_len = pw_len + 16;
u32 w0_t[4]; u32 w0_t[4];
u32 w1_t[4]; u32 w1_t[4];
u32 w2_t[4]; u32 w2_t[4];
@ -551,21 +561,19 @@ static void m09800s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
u32 key[4]; digest[0] = swap32 (digest[0]);
digest[1] = swap32 (digest[1]);
key[0] = swap32 (digest[0]); digest[2] = swap32 (digest[2]);
key[1] = swap32 (digest[1]); digest[3] = swap32 (digest[3]);
key[2] = swap32 (digest[2]);
key[3] = swap32 (digest[3]);
if (version == 3) if (version == 3)
{ {
key[1] &= 0xff; digest[1] &= 0xff;
key[2] = 0; digest[2] = 0;
key[3] = 0; digest[3] = 0;
} }
rc4_init_16 (rc4_key, key); rc4_init_16 (rc4_key, digest);
u32 out[4]; u32 out[4];

21
src/shared.c
View File

@ -9115,27 +9115,6 @@ void read_restore (const char *eff_restore_file, restore_data_t *rd)
fclose (fp); fclose (fp);
char new_cwd[256] = { 0 };
char *nwd = getcwd (new_cwd, sizeof (new_cwd));
if (nwd == NULL)
{
log_error ("Restore file is corrupted");
}
if (strncmp (new_cwd, rd->cwd, sizeof (new_cwd)) != 0)
{
if (getcwd (rd->cwd, sizeof (rd->cwd)) == NULL)
{
log_error ("ERROR: could not determine current user path: %s", strerror (errno));
exit (-1);
}
log_info ("WARNING: Found old restore file, updating path to %s...", new_cwd);
}
if (chdir (rd->cwd)) if (chdir (rd->cwd))
{ {
log_error ("ERROR: cannot chdir to %s: %s", rd->cwd, strerror (errno)); log_error ("ERROR: cannot chdir to %s: %s", rd->cwd, strerror (errno));